Movienet: Deep space–time‐coil reconstruction network without k‐space data consistency for fast motion‐resolved 4D MRI

Purpose To develop a novel deep learning approach for 4D‐MRI reconstruction, named Movienet, which exploits space–time‐coil correlations and motion preservation instead of k‐space data consistency, to accelerate the acquisition of golden‐angle radial data and enable subsecond reconstruction times in dynamic MRI. Methods Movienet uses a U‐net architecture with modified residual learning blocks that operate entirely in the image domain to remove aliasing artifacts and reconstruct an unaliased motion‐resolved 4D image. Motion preservation is enforced by sorting the input image and reference for training in a linear motion order from expiration to inspiration. The input image was collected with a lower scan time than the reference XD‐GRASP image used for training. Movienet is demonstrated for motion‐resolved 4D MRI and motion‐resistant 3D MRI of abdominal tumors on a therapeutic 1.5T MR‐Linac (1.5‐fold acquisition acceleration) and diagnostic 3T MRI scanners (2‐fold and 2.25‐fold acquisition acceleration for 4D and 3D, respectively). Image quality was evaluated quantitatively and qualitatively by expert clinical readers. Results The reconstruction time of Movienet was 0.69 s (4 motion states) and 0.75 s (10 motion states), which is substantially lower than iterative XD‐GRASP and unrolled reconstruction networks. Movienet enables faster acquisition than XD‐GRASP with similar overall image quality and improved suppression of streaking artifacts. Conclusion Movienet accelerates data acquisition with respect to compressed sensing and reconstructs 4D images in less than 1 s, which would enable an efficient implementation of 4D MRI in a clinical setting for fast motion‐resistant 3D anatomical imaging or motion‐resolved 4D imaging.